How to build an eight-week AI training program for your team

Training a team in AI is not running a workshop: it means sustaining eight weeks of real work with a diagnostic, a rhythm, internal champions and measurement. Here is the complete program, with the agenda for every session, the exercises by department, the budget, and the tips that only show up once you have run it several times.

Why eight weeks and not a one-day workshop?

Because one day creates enthusiasm and eight weeks create a habit. A standalone workshop works as a demo: people walk out impressed, open the tool twice that afternoon, and ten days later they are working exactly as before. Eight weeks is the minimum for a team to go through two real work cycles —a month-end close, a two-week collections run, a proposal cycle— using the tool inside those cycles and not in a lab exercise.

There is a second, more uncomfortable reason: the first month of use usually produces mediocre results. People ask badly, accept the first answer, skip verification, and conclude that "it doesn't work for what I do". If the program ends in week two, that conclusion sticks. If it continues, by week four that same person is providing context, has a project with their own documents inside it, and knows when to be skeptical. The curve dips before it rises.

Anthropic's free official training framework is called AI Fluency and organizes the skill set into four dimensions —delegation, description, discernment and diligence— in a course of about four hours (Claude Academy, accessed September 2026). It is an excellent foundation, but it is not a company program: it is missing the part you have to supply, which is your people's real work.

A calendar note for Mexico and Paraguay: pick eight weeks that don't collide with the fiscal close, your sector's high season, or your industry's big trade show. A program that kicks off in October at a company exhibiting in November is doomed before it starts.

How do you measure the team's real level before you start?

With a short behavioral test, not a knowledge test. The question "do you know how to use AI?" is useless: almost everyone says yes and almost nobody means the same thing. What you need to know is what the person actually does when they sit down in front of the tool, how often, and how far they get. Ten questions are enough and take four minutes to answer.

Two warnings. Question 6 is the most revealing: if more than half answer "I take it at face value", start with verification and not with productivity. And keep the responses, because without a baseline there is no defensible report afterward.

Almost always the same finding shows up: question 1 returns higher numbers than management expects. People are already using AI on personal accounts with no rules. That is a data point, not an offense, and the place to start is the post in this series on an internal AI use policy.

What participant profiles will you run into?

Six, almost always the same ones. Recognizing them in the first session saves weeks, because each one drops out for a different reason and gets hooked by a different one.

The practical consequence: you cannot run one identical session for everyone. And the cheap way to solve it is not splitting the program into two levels, but splitting the practice: 45 minutes of shared live content and guided exercises at tables of three or four people from the same department, each with a different brief.

The eight-week calendar, week by week

Each week has a single objective, 45 to 60 minutes of live content, guided practice and a small deliverable. The rule that holds it all together: the deliverable is always the participant's real work. If the week doesn't produce something the person was going to do anyway, it is badly designed.

Week 1 — What it is, what it isn't and how you talk to it

Objective: everyone finishes with a real task solved and a correct idea of what tool they have in front of them. Content: how a language model works in useful terms (it predicts text, it doesn't query a database), what context is and why it sometimes makes things up. Practice: rewriting a real, difficult email and comparing the one-line version with the version that includes context. Support: the "How AI Works" series from Claude Academy.

Week 2 — Describing the task well

Objective: move from asking to assigning. Content: the golden rule from the official documentation —"show your instruction to a colleague with little context; if they get confused, so will the model"— and the three levers with the highest payoff: give the reason, include three to five examples, and separate instructions, context and data with tags (official prompting guide, September 2026). Deliverable: a saved, reusable instruction. If the team is a sales team, the post on prompts for sales teams is the week's textbook.

Week 3 — Working on your own documents

Objective: get people out of "plain chat". Content: uploading files, asking for extraction and comparison, working with tables, generating output documents. Tip from the documentation: when you paste a long text, put it at the top and the question at the bottom; in Anthropic's testing that improves quality by up to 30%. Deliverable: a comparison of two real documents from the department.

Week 4 — The project: your department's memory

Objective: every department gets its own space with its own knowledge inside. Content: what goes into the knowledge base, how project instructions are written, and who can see it. On paid plans, projects use retrieval augmentation, which according to the documentation expands capacity "up to 10 times" as you approach the context limit (Projects documentation). Deliverable: one project per department with five documents and its instructions written out. This is the hinge of the program, and it is developed in full in the post on Projects as your company knowledge base.

Week 5 — Judgment: verify, doubt and leave a record

Objective: install the reflex of checking. Content: what kinds of mistakes a model makes (plausible figures, references that don't exist, excessive confidence on ambiguous topics), how to ask for citations before the conclusion, and what always gets verified: numbers, proper names, dates, regulatory references and anything that leaves the company. Deliverable: the department's verification rule, in five lines. A reminder worth repeating: nothing that comes out of an AI tool replaces professional legal, tax or accounting advice.

Week 6 — Connecting it to what you already use

Objective: take copy-paste out of the middle. Content: connectors for email, calendar, storage or CRM, and the permission decisions behind them (connectors documentation). Deliverable: one weekly departmental workflow, step by step, with the exact point where the tool comes in. If it starts to smell like real automation, it is already a process automation project, not a training exercise.

Week 7 — Reusable templates

Objective: make sure the knowledge doesn't live inside individual heads. Skills —folders of instructions and resources the model loads when they are needed— are available on all plans, are written in Markdown, and on Team and Enterprise the owner can provision them across the whole organization (skills documentation). Deliverable: two departmental templates published where everyone can see them.

Week 8 — Internal demo and 90-day plan

Objective: close with evidence. Each department presents a five-minute before-and-after with its own numbers, and leaves behind one slide and three 90-day commitments. Invite management to this session and only this one: it is the one that decides whether there is a second cohort.

With fewer than twelve people you can compress it to six weeks by merging weeks 3 and 4 and weeks 6 and 7. Below six weeks it is no longer a program, it is a long workshop.

Which hands-on exercises should each department get?

The ones that produce something that was going to be done anyway. This table is a starting point for mid-sized companies in Mexico and Paraguay; adapt it to your operation, not the other way around.

Two rules for building your own. The exercise has to fit in 25 minutes, because otherwise it doesn't get done in the session and it doesn't get done afterward either. And always ask for the before: have the person bring the version they made by hand last month. Without that before, week eight has nothing to compare against. There is a broader catalog in the post on use cases by department.

Which format works: live session, guided practice or weekly challenge?

All four, in very uneven proportions. The combination that survives eight weeks at a small or mid-sized company is: 60 minutes live, 30 minutes of guided practice inside the same block, a 20-minute weekly challenge outside the session, and a permanent internal channel. Everything else is decoration.

On the channel: use the tool your company already has open all day. At many small and mid-sized companies in Mexico and Paraguay that means a WhatsApp group, and that's fine: a group people actually read works better than a corporate forum nobody opens. Set only two things: what does not get pasted there (customer data, credentials, confidential documents) and who answers. If you end up providing systematic support through that channel, turn it into a process with WhatsApp Business with AI.

On timing: the best slot in an office is first thing Tuesday or Wednesday, because Monday fills up with emergencies and Thursday and Friday with closings. With shift work or people on the road, run the session twice at two different times instead of recording it.

Which free official materials can you use at no cost?

Quite a few, and of better quality than most of the paid material out there. Anthropic's official training is at Claude Academy and, as of September 2026, every course listed is free.

• AI Fluency: Framework and Foundations — 14 lessons, about 4 hours. The backbone of weeks 1, 2 and 5.
• AI Capabilities and Limitations — 13 lessons, about 3.5 hours. Perfect material for the judgment week.
• Building Effective Human-Agent Teams (Beta) — 5 lessons, 45 minutes. For when the team moves from individual use to shared work.
• Claude 101, with the Introduction to Projects module, for week 4.
• Short tutorials at academy.claude.com/tutorials, 3 to 20 minutes each. The highest-return ones: the Enterprise administrator guide (20 min, for IT), "How context affects Claude's performance and cost" (20 min) and "Writing an AI diligence statement" (15 min).
• Prompting documentation — the prompt engineering section, the source of the week 2 rules.
• Interactive tutorial — public GitHub repository, ideal for the technical profile, and the Anthropic Discord for your champions.

Also keep the model comparison table handy: it settles half a dozen recurring questions about when a fast model is the right call and when you want a more powerful one.

How does the internal champion dynamic work?

A champion is the person in a department everyone asks without booking a meeting. Not the most technical one and not the boss: the one who already had a reputation for helping. With one champion for every eight to twelve people, the program stops depending on the trainer right around week four, which is exactly when the trainer becomes the bottleneck.

How to pick them. Not by fiat over email. Ask in session 2: "who in your department do you go to when something on your computer gets stuck?". The same names come up on their own. Then talk to each of them privately and offer it as an opportunity, not a burden.

What they do. Three things and nothing else: answer questions from their department in the channel, bring a real case to every session, and collect their people's deliverables. They don't teach, they don't assess, and they don't report who is using it and who isn't. The moment a champion turns into an undercover supervisor, people stop asking them anything.

What they get. A 45-minute prep session each week, an explicit mention from management in the final session, and above all recognized time: two or three hours a week blocked off with their direct manager's sign-off. Without that recognition, a champion lasts three weeks.

A side effect worth anticipating: champions end up knowing more than their bosses. If your company culture handles that badly, talk to the department heads first and give them the role of visible sponsor.

How do you keep people from dropping out in week three?

By anticipating the five real causes of attrition. They are always the same and none of them has to do with interest.

• The challenge doesn't look like their job. If the exercise is "summarize this sample article", the person does the math correctly: this saves me nothing. Phrase every challenge with a verb from the role ("prepare the proposal you were going to send on Thursday") and let them pick the document.
• They don't have access. More common than it sounds: the license was approved but never provisioned, or the person doesn't know how to log in. Check access one by one before session 1 and again in week 3. Nobody raises their hand to say "I haven't been able to log in for two weeks".
• The session collides with production. A badly chosen time slot generates absences that get read as disinterest. If two people from the same shift miss two weeks in a row, the problem is the schedule.
• Unnamed fear. As long as nobody says out loud what this is for, part of the team assumes the worst. Management has to say it in session 1 in one short sentence, and that sentence has to be true.
• Nothing happens if you don't submit. You don't need punishment, you need visibility: a simple board showing who submitted, read out in thirty seconds at the start of each session.

Three cheap tactics that work: pair each participant with someone from another department for the challenge, close every session with two minutes of "this helped me this week" from a different person each time, and send the reminder on the same day and at the same time, always. Consistency matters more than the wording.

A realistic expectation, from classroom experience rather than published statistics: live attendance drops somewhat toward week four and levels off there. That is normal. What should worry you is the challenge submission rate: if it falls below 50%, stop and review the brief before you go on.

How do you assess progress without inventing certifications?

With work evidence and a short rubric, not with exams. And with the honesty rule already stated: there is no official Anthropic certification, so any recognition you issue has to say clearly that it is internal.

The assessment is done by the direct manager, not the trainer: they look at the submissions and mark the level, ten minutes per person. If you want a symbolic closing, issue a certificate with text along the lines of "internal recognition of participation in the applied AI training program, 2026 edition, granted by [your company]". Never name Anthropic or a vendor as the issuer, because they are not.

For the report to management the rubric boils down to one line: how many people are at level 3 or above at the end of week eight, compared with the initial diagnostic. The financial return is a separate matter with its own method, laid out in full in the post on how to measure adoption and return.

What you should not do: multiple-choice exams, public rankings of individuals, and measuring by number of messages sent. The last one is the most tempting and the worst: it rewards whoever types the most, not whoever works best.

Agenda template for a 90-minute session

The same agenda for all eight weeks. The repetition is part of the method: people know what is going to happen and stop spending energy adapting to the format.

A detail that saves a lot of time: prepare the guided practice as one document per department with the brief, the sample file and the "done" criterion. If the table has to decide what to do, you lose ten of your thirty minutes.

What does it cost to run the program?

Less than people assume in licenses and more than people assume in hours. The dominant cost is always people's time.

Licenses, with prices as of September 2026, in dollars, before tax, and worth confirming on their site because they vary by region: Pro costs USD 20 per month (17 with annual billing); Team, USD 25 per seat per month (20 annual), with a minimum of 2 and a maximum of 150 seats; self-serve Enterprise starts at USD 20 per seat per month plus usage at API rates, with a minimum of 20 seats (claude.com/pricing and Team plan documentation).

There is a non-financial reason to choose a commercial plan that weighs more than the price: in the commercial products, by default Anthropic does not use your inputs or outputs to train its models (privacy documentation, September 2026). Training thirty people to work with company documents on personal accounts is a governance problem you pay for later.

The figures in the third column are an illustrative example, not a price list. Swap in your own hourly cost (gross salary plus payroll burden divided by working hours) and the conversation with management stops revolving around licenses in under a minute: in the example, those 596 hours are worth many times what the seats cost.

Before you finalize the budget, look into local funding: in Mexico, development banking includes business training among its services —check current terms at Nacional Financiera and ask your chamber of commerce, for example CANACINTRA—; in Paraguay, trade associations such as the Unión Industrial Paraguaya organize sector training. None of them publishes a specific AI program with fixed terms, so ask before you count it as income.

How do you get management's backing?

By asking for three specific things and offering one. What you ask for: a budget for eight weeks, protected time on people's calendars, and visible sponsorship. What you offer: a report with numbers in week eight, using a method agreed in advance. That last point changes the conversation, because it prevents someone in week eight from asking you for a figure you never recorded.

On sponsorship: five minutes from management in the first session and their presence at the last one is enough, but it has to be real; a generic video recorded once conveys the opposite of what it intends. And if the sponsor also does the week 1 challenge and talks about it, that convinces a skeptic faster than any speech.

If someone asks for a comparable case before approving, the strongest argument is not somebody else's case but a pilot of your own: two departments, three weeks, with measurement. If it moves nothing, there is no program and you save yourself the big budget. That phased approach is the one in the pillar post on how to roll out Claude at your company.

Tips that don't come in the manual

This is what you learn from running the program several times and what appears in no product guide.

1. Start with the task people hate, not the most profitable one. The first case should be the tedious work nobody defends: transcribing notes, reconciling formats, writing up the minutes. Adoption is emotional before it is economic; the profitable task arrives in week four.
2. Ban the word "prompt" for the first two weeks. Say "assignment" or "instruction". The jargon triggers the "this is technical and not for me" reflex in exactly the profile that is hardest to hook.
3. Long document at the top, question at the bottom. When you paste a contract or a long report, the text goes first and the question at the end. The official documentation notes that putting the query at the end can improve quality by up to 30% in its testing. It is the best effort-to-result trick in the program.
4. Ask for the citations before the conclusion. "First pull the verbatim sentences from the document that support the answer; then answer." It reduces invention and leaves your verification half done.
5. Three to five examples, not one. The official guide recommends between three and five, diverse and including edge cases. A single example produces clones of that example.
6. Explain the reason, not just the rule. Instead of "don't use jargon", write "a customer with no technical background is going to read this, so avoid jargon". The example is in the documentation itself, and the difference in output is obvious.
7. The colleague test. Before you consider an instruction good, show it to someone with little context. If they get confused, so will the model. It's the golden rule from the official guide and it works as free quality control.
8. Turn every question in the channel into material. If a question comes up twice, it is next week's case. In eight weeks the program writes itself from your people's real questions, and that material serves the second cohort.
9. Ask for the "before" in session 1. Have each participant bring the last version they made by hand and how long it took them, written down at that moment and not reconstructed in week eight. Memory about time spent is catastrophically optimistic.
10. Keep the bad answers in a shared document. A "what went wrong" file with screenshots and explanations. It's the best material for week 5, it wins over the skeptic, and it shows management that the team has judgment, not faith.
11. A template is not a text, it's a folder. In week 7, put the sample file and the "done" criterion next to the instruction. An instruction without an example degrades within three uses.
12. Close with a demo, not a survey. Five minutes per department showing something that works, in front of management. A satisfaction survey does not get you the second cohort; a demo with two numbers does.

What to do this week

If you're about to start, this is the order that avoids the classic stumbles. It all fits in a normal week.

1. Monday. Send the 10-question test to candidate participants, with a 48-hour deadline.
2. Monday. Ask IT for the list of who has a license today and what type. If personal accounts are in use, note it without making a drama of it.
3. Tuesday. Pick two or three departments for the first cohort. Prioritize repetitive volume and a willing manager, not enthusiasm.
4. Wednesday. Write the one-page memo using the template above, with a specific amount and specific dates.
5. Wednesday. Record the baseline: two measurable tasks per department, with current time and monthly volume.
6. Thursday. Book all eight sessions, same day and same time, with the guided practice in the same block.
7. Thursday. Identify the champions by asking who people go to, and talk to each of them privately.
8. Friday. Verify access one by one: nobody should arrive at session 1 unable to log in.
9. Friday. Open the internal channel with the rules on what doesn't get shared there, and post the session 1 agenda.
10. Friday. Prepare the session 1 case with a real, anonymized company document. No textbook examples.

If you'd rather not carry it alone, outside support pays off above all in designing the baseline and in weeks 4 and 7, the most demanding ones; at Uniamos we handle this within AI training with Claude, and you can stress-test your plan by writing to us through contact. One last warning: training on top of a broken process only speeds up the mess, so if the problem is the process, fix that first.

Frequently asked questions

Related articles

• How to roll out Claude at your company in phases: 2026 guide
• How to measure AI training adoption and return
• Projects: your company's knowledge base
• Claude prompts and tips for sales teams
• Internal AI use policy and governance
• Claude use cases by department in SMBs
• AI training with Claude for teams
• More from the blog